Drop an exam PDF in. Get a timed, real-feeling computer test: split-screen with the source document, question palette, flagging, auto-marking, and a downloadable results PDF. Runs 100% on your machine — no account, no server, no cloud.
exam PDF ──▶ TestPractice ──▶ timed split-screen exam ──▶ auto-marked results + printable PDF
(optional: your AI builds the test via local MCP)
- Add a document — PDF, Word, RTF, EPUB or CSV. Text-based files parse instantly; scanned ones are OCR'd locally with Tesseract (~20 s for a 20-page scan).
- Build the test — three ways:
- Instant: text-based PDFs are extracted automatically.
- Ask your AI (recommended): your own ChatGPT / Claude / Codex connects to the
local MCP endpoint (
http://localhost:5874/mcp, no installation) and writes the test JSON. A ready-made prompt is on the landing page. - JSON import: bring your own file; the validator checks it before the exam.
- Set up the attempt — override the timer (custom minutes or untimed free practice) and pick which questions you want: quick-select range blocks (Q1–10, Q11–20 …) let you drill one chapter or section without touching the JSON.
- Take the exam — countdown timer with auto-submit, question palette (○ unanswered ● answered ⚑ flagged ◉ current), split-screen passage view, autosave + resume, review-before-submit.
- Results — auto-marking with per-question comparisons (right / wrong / manual), downloadable PDF summary.
Question types: single_choice, multiple_choice, true_false (T/F/N),
text_input, matching, long_text. Tests are plain JSON files — portable,
versionable, human- and AI-writable.
Download one file for your platform from Releases:
| Platform | File |
|---|---|
| Windows 10/11 x64 | TestPractice-windows-x64.zip → TestPractice.exe |
| Linux x64 | TestPractice-linux-x64.zip → TestPractice |
| Apple Silicon (M1–M3) | TestPractice-macos-arm64.zip → TestPractice |
| macOS Intel (2014+) | see PLATFORMS.md — build locally on a Mac |
Unzip, double-click. A tray icon appears (open in browser / quit); the browser
opens automatically at http://localhost:5874. Everything — server, UI, PDF
renderer, parser, report writer — is inside the binary; tests/ and pdfs/
folders are created next to it on first run.
Optional: install Tesseract for local OCR of scanned PDFs (apt / brew / UB-Mannheim installer). Without it the app still works — it just asks your AI to read the rendered pages instead.
No framework, no build step:
python3 server.py # serves http://localhost:5874 (MCP at /mcp)| File | Role |
|---|---|
server.py |
stdlib HTTP server: static files, /upload-doc, /export-results, MCP JSON-RPC (list_tests get_test save_test validate_test list_pdfs parse_pdf prepare_test_from_pdf) |
launch.py |
launcher: starts the server, opens the browser, tray icon to quit |
index.html app.js style.css pdf.js |
the app — vanilla JS, PDF.js vendored (no CDN) |
tests/*.json |
test files (schema validated by validate_test server-side and validateTest in app.js) |
build.sh / build.bat |
per-platform frozen builds (GitHub Actions does this per tag too) |
Security posture: binds to 127.0.0.1 only; filename validation on every route;
no telemetry, no accounts, fully offline (PDF.js is vendored — zero outbound calls).
Apache-2.0. Third-party: PDF.js (Apache-2.0), firecrawl-anydoc (MIT), reportlab (BSD), PyInstaller (GPL with the bootloader exception covering the bundled output), pystray (LGPL-3.0, dynamically linked), Pillow (MIT-CMU), Tesseract (Apache-2.0).